Dynamically updating parameter each iteration of NDSolve

I am struggling to update the value of a parameter inside NDSolve that changes with the current value of my state. See an extremely simplified version of the code below.

x[t] = {x1[t], x2[t]};

x'[t] = D[x[t], t];

Cmat = {{10, x1[t]}, {x2[t], 30}};

Cmatinv = Inverse[Cmat /. x1[t] -> 3 /. x2[t] -> 5]

{{2/19, -(1/95)}, {-(1/57), 2/57}}

NDSolve[{x'[t] == Cmatinv. x[t], x1 == 3, x2 == 5},
x[t], {t, 0, 5}];

Essentially what I am trying to do is recalculate Cmatinv at each iteration within NDSolve. The problem is that, in my real code, Cmat is a 20x20 with large expressions for each element, so I can't do a symbolic inverse like I could for the simplified code above. To initialize the differential equation, I substitute the initial conditions into Cmat so that the inverse is done on a numeric matrix and is therefore extremely fast. I can't figure out how to then update this at each iteration with the new states within NDSolve. The procedure I want to accomplish within each NDSolve iteration is: calculate x[t], calculate Cmat (numeric), calculate Cmatinv, solve differential equation. I've been troubleshooting this problem for months, so any help would be appreciated. I've seen similar questions answered, but those evaluate the relevant parameters as functions of the state. As far as I can tell, I can't do this, because the inverse would end up being symbolic.This is my first time posting, so I apologize if I'm missing any rules/information.

I am including a more detailed version of my code, because the simplified version doesn't really show exactly what I'm struggling with.

(*Physical Parameters*)
param1 = 8.99*10^9;
param2 = 9.81;

(*System Parameters*)
n = 20;
tmax = 0.1;

(*Spring Coefficient*)
springparam = 2.9*10^-6;

(*Material Properties and Mass Calculation*)
length = 0.06527392544827149;
l = length/n;
m = 8.668377299530452*10^-8;

(*Field Calculations*)
sysparam = 6150;
sphparam = ConstantArray[sysparam, n];
fieldparam = {60531.496062992126, 0, 0};

(*System Setup*)
h = {0, 0, 0};
h[i_] := h[i - 1] +
l {Sin[Subscript[θ, i - 1][t]], -Cos[Subscript[θ, i - 1][t]], 0}
hinge = Table[h[ii], {ii, n}];
com = Table[
hinge[[i]] + l/2 {Sin[Subscript[θ, i][t]], -Cos[Subscript[θ, i][t]], 0}
, {i, n}];
vcom = Table[D[com[[i]], t], {i, n}];
comypos = Table[-(2 i - 1) l/2, {i, n}];

initpos = Table[Subscript[θ, i][t] -> 0, {i, n}];

r[i_, j_] := Sqrt[(com[[i, 1]] - com[[j, 1]])^2 + (com[[i, 2]] - com[[j, 2]])^2]
r[i_, i_] := 0.0009242787843475243;
invmat = (Table[(r[i, j])^-1, {i, n}, {j, n}]);

mat = 1/param1 Inverse[invmat /. initpos];
val = mat.sphparam;

(*Generalized Force Setup*)
forceparam = Table[fieldparam*val[[i]], {i, n}];
genforceparam = Table[Sum[
forceparam[[k]].D[com[[k]], Subscript[θ, i][t]]
, {k, n}], {i, n}];

(*Langrangian Setup*)
u = Table[
m param2 (com[[i, 2]] - comypos[[i]]) +
1/2 springparam*(Subscript[θ, i][t] - Subscript[θ, i - 1][t])^2, {i, 1, n}];
u[] = m param2 (com[[1, 2]] - comypos[])
+ 1/2 springparam (Subscript[θ, 1][t] - π/2)^2;
k = Table[1/2 m vcom[[i]].vcom[[i]] + 1/24 m l^2 Subscript[θ, i]'[t]^2, {i, n}];

L = Sum [k[[i]], {i, n}] - Sum[u[[i]], {i, n}];
eulerlagrange[i_] :=
D[D[L, Subscript[θ, i]'[t]], t] - D[L, Subscript[θ, i][t]] == genforceparam[[i]];
diffeq = Join[Table[eulerlagrange[i], {i, n}],
Table[Subscript[θ, i] == 0, {i, n}],
Table[Subscript[θ, i]' == 0, {i, n}]];

evalel =
NDSolveValue[diffeq,
Join[Table[Subscript[θ, i][t], {i, n}],
Table[Subscript[θ, i]'[t], {i, n}]], {t, 10^-6, tmax },
Method -> {"EquationSimplification" -> "Residual"}]

You can see here that, in order to get my generalized force for the Euler-Lagrange equation, I have to solve for val, meaning that I need to invert invmat which depends on com.

• Is there a reason why you have to multiply with the inverse on the right hand side rather than just multiplying with the original matrix on the left hand side? – Sjoerd Smit Jan 17 '18 at 14:59
• I tried @SjoerdSmit's idea to put the original matrix on the left hand side of NDSolve. It seems to work fine. – Chris K Jan 17 '18 at 18:38
• I tried your second example. It gave a warning NDSolveValue::ivres, but otherwise produced a solution. Could you pinpoint the problem with this code? – Chris K Jan 18 '18 at 21:28
• Chris K, it will run, but the invmat, mat, val, etc. should change on every iteration of the loop, but they don't (as far as I know). The obvious way to do this is to make it a function, but if I do, the inverse evaluates symbolically, which takes forever for a 20x20. What I'm really trying to do is figure out how to tell mathematica that invmat MUST be populated with numbers before the inverse is evaluated on each iteration of NDSolve. – Jordan Maxwell Jan 19 '18 at 16:33

Since NDSolveValue works with vectors, I would tweak @HenrikSchumacher's answer as follows:

inv[x_List] := LinearSolve[{{10, x[]}, {x[], 30}}, x]

sol = NDSolveValue[
{x'[t] == inv[x[t]], x == {3, 5}},
x,
{t, 0, 5}
];

Here are a couple plots:

Plot[sol[t], {t, 0, 5}] ParametricPlot3D[{t, sol[t][], sol[t][]}, {t, 0, 5}] • Ah, very good; it does work. Mathematica complained about a list element that cannot be accessed, so I refrained from this. Using inv[x_List] := ... is the fix. – Henrik Schumacher Jan 17 '18 at 17:20
• Thanks all for the feedback. In looking over the results, I think the example I provided was a bit too simple. I've added a more-detailed version of the code for clarity. – Jordan Maxwell Jan 18 '18 at 16:11

ClearAll[x, y];
F[t_, x_, y_] := LinearSolve[{{10., x}, {y, 30.}}, {x, y}];
{x, y} = NDSolveValue[
{
{x'[t], y'[t]} == F[t, x[t], y[t]],
x == 3,
y == 5
},
{x, y},
{t, 0, 5}
]
ParametricPlot3D[{t, x[t], y[t]}, {t, 0., 5.}] Side remark: Avoid Inverse for all matrices bigger than, say $6 \times 6$.

To follow-up on my comment above, on @SjoerdSmit's observation that you could multiply the left hand side by the original matrix to avoid inverting it:

sol = NDSolve[{{{10, x1[t]}, {x2[t], 30}}.{x1'[t], x2'[t]} == {x1[t], x2[t]},
x1 == 3, x2 == 5}, {x1, x2}, {t, 0, 5}];
Plot[{x1[t], x2[t]} /. sol, {t, 0, 5}] Looks like the other answers above.